Skip to content

Retire lightpanda from the profiles; ego-browser takes the slot - #129

Merged
NagyVikt merged 3 commits into
mainfrom
chore/retire-lightpanda-profiles
Aug 7, 2026
Merged

Retire lightpanda from the profiles; ego-browser takes the slot#129
NagyVikt merged 3 commits into
mainfrom
chore/retire-lightpanda-profiles

Conversation

@NagyVikt

@NagyVikt NagyVikt commented Aug 7, 2026

Copy link
Copy Markdown
Contributor

Follow-up to #126, which merged before these commits landed on it — it carried
only the submodule bump, so main still declares a skill that no longer exists
upstream.

What's wrong on main right now

opencue/skills#20 deleted browser/lightpanda. main's core profile still
declares it, and main's skills pin (ba3d06d) predates the deletion. The two
agree with each other, so nothing is broken — they are just both stale, and any
checkout that advances the submodule gets SKILL_NOT_FOUND.

Why lightpanda is going

#121 removed the lightpanda MCP from every profile that pinned it, but left
the skill. Of the 29 skill→MCP requirements in the tree, browser/lightpanda
was the only one with no provider at all
— it could not run, yet core shipped
it to every session and the smart-loader kept offering it, annotated (needs MCP: lightpanda).

ego-browser covers the same ground — headless navigation, scraping, DOM
extraction — and also paints pixels, which was the gap that forced the old
two-engine handoff in the first place.

Changes

  • core and browser no longer declare browser/lightpanda.
  • Four personas that routed agents to it now route to ego-browser: browser
    (its "default browser engine" section), commerce and webshop (their "which
    browser when" tables), plus stale MCP-inheritance comments in gstack and
    x-growth-bot. agent-browser and webwright keep their entries — the
    persistent Rust daemon and the long-horizon mission workflow are still theirs.
  • resources/skills752aef4 (Retire the lightpanda skill skills#20: the deletion plus the four
    in-repo references it left dangling, including verify-council's visual-proof
    lane, which had been told to drive a dead MCP).

Note on summon.test.ts

I had rewritten it here against a hardcoded deployment/coolify + backend
pair. #127 landed a better fix on main — deriving a still-valid pairing
from live profile data via findMcpGatedPair() — so I dropped mine and kept
main's during the cherry-pick. Nothing from this branch touches that file.

Verification

Full suite on this branch, with the skill gone from disk and from every profile:
3077 pass, 0 fail (3078 across 231 files, 1 skip).

grep -rn lightpanda profiles/ returns one deliberate line — a note in gstack
recording that core used to ship the MCP, so the next reader knows it went away
rather than never existed.

NagyVikt and others added 3 commits August 7, 2026 06:11
so `core` shipped a skill that asks for a server nothing provides. Every session
carried it, and the smart-loader kept offering it with "needs MCP: lightpanda".
ego-browser covers the same ground — headless navigation, scraping, DOM
extraction — and it also paints pixels, which was lightpanda's stated gap.

- `core` and `browser` no longer declare `browser/lightpanda`.
- The `browser` persona named Lightpanda as its default engine and explained how
  to fall back to Chromium for screenshots. Rewritten around ego-browser, with
  the playwright MCP kept for the case ego-browser genuinely doesn't cover:
  Firefox and WebKit.

`summon.test.ts` keyed its mcp_status assertion off browser/lightpanda + core.
That pair can no longer reach the "ok" branch, and it only passed this long
because the committed catalog index was stale and still named core as a
lightpanda provider. Rebuilt on deployment/coolify + backend: the summoned
profile ships the skill, a different profile supplies the MCP, which is the
distinction the test exists to make. Comment left in place so nobody restores
the old pair.

Full suite: 3023 pass, 0 fail.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Four personas still told agents to reach for lightpanda, which no profile can
supply any more — commerce and webshop each carried a "which browser when"
section naming it as the scrape/read engine, and gstack and x-growth-bot both
described core as shipping the lightpanda MCP.

ego-browser takes that slot. The old split existed because lightpanda painted no
pixels, so the guidance had to hand off to something else the moment a task
needed a screenshot; ego-browser does both, so the rule of thumb collapses to one
engine. agent-browser and webwright keep their entries — the persistent Rust
daemon and the long-horizon mission workflow are still theirs — and the tables
still list three engines, just with the first one swapped.

Left one deliberate mention: the gstack comment now records that core *used to*
ship lightpanda, so the next person to read it knows the MCP went away rather
than wondering if it was never there.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
opencue/skills#20 removes the browser/lightpanda skill and repoints the four
remaining references (verify-council's visual-proof lane, agent-browser's
overlap note, cue-dashboard's sample output) at ego-browser. Catalog 454 -> 453.

Full suite with the skill fully gone from disk: 3023 pass, 0 fail.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
@NagyVikt

NagyVikt commented Aug 7, 2026

Copy link
Copy Markdown
Contributor Author

Automated review — verdict: no blocking findings. 0 CRITICAL, 0 HIGH, 0 MEDIUM, 0 LOW.

Three things worth checking on a change that deletes a skill out from under a
profile; all three hold.

Nothing else references the skill. grep -rn lightpanda profiles/ src/ is
clean apart from one deliberate note in gstack recording that core used to
ship the MCP. The four in-repo references inside the skills submodule
(verify-council's visual-proof lane, agent-browser's overlap note,
cue-dashboard's sample output) were repointed in opencue/skills#20, which this
bump pulls in — so no dangling instruction survives on either side.

The persona claims about ego-browser are accurate, not aspirational. The
rewritten sections say it paints pixels (page.screenshot()), that it is
Chromium-only, and that playwright stays for Firefox/WebKit. Checked against the
skill's own SKILL.md rather than asserted.

The bump does not lose anything. resources/skills moves to 752aef4; the
only removal in that range is browser/lightpanda itself, and ego-browser,
agent-browser and playwright all survive.

Verification

Check Result
CI — test Pass — and main's latest CI run is green too, so this clears an actually-green baseline, not a red one
CI — lint, configured Pass
CI — Profiles e2e (ubuntu + macos) Pass
CI — validate Fail, pre-existing and unrelated: E3 (missing reference): MCP "openaiDeveloperDocs", and nothing else. profiles/research declares that MCP, it is registered in no config in the mcps repo, and it is not installed on this machine. #121 failed the same check and merged; the job is pull_request-only so it has no main baseline
Local full suite 3077 pass, 0 fail (3078 across 231 files, 1 skip)

Note

I had also rewritten summon.test.ts here against a hardcoded
deployment/coolify + backend pair. #127 landed a better fix on main
deriving a still-valid pairing from live profile data via findMcpGatedPair()
so mine was dropped during the cherry-pick. This branch does not touch that file.

@NagyVikt
NagyVikt merged commit 4ccb929 into main Aug 7, 2026
6 of 7 checks passed
@NagyVikt
NagyVikt deleted the chore/retire-lightpanda-profiles branch August 7, 2026 04:15
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant